Pith. sign in

REVIEW 5 major objections 6 minor 1 cited by

Towards White-Box Deep Wireless Sensing

T0 review · 5 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read RF-CRATE shows that a fully complex-valued transformer for wireless sensing can be derived end to end from the sparse rate reduction principle, and the derived model matches black-box accuracy on five datasets.

desk verdict Solid empirical package with a mismatched title: the model is a complex-valued white-box-style transformer, but the derivation has hand-inserted steps, so the 'fully derived' claim is ahead of the evidence. read the letter →

arxiv 2507.21799 v2 pith:2IQT36IF submitted 2025-07-29 cs.LG cs.AI

Xie Zhang , Yina Wang , Chenshu Wu This is my paper
classification cs.LGcs.AI
keywords wirelesssensingRFwhite-boxtransformercomplex-valuedneuralnetworkssparseratereductionalgorithmunrollingsubspaceregularizationinterpretability
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

RF-CRATE aims to show that a deep network for wireless sensing can be built from a stated mathematical principle rather than assembled from heuristics. The paper derives a fully complex-valued transformer whose self-attention and residual MLP modules come from optimizing a complex sparse rate reduction objective, using the CR-Calculus framework to handle the fact that the objective is not holomorphic. The model is tested on WiFi, mmWave radar, and UWB data across gesture, gait, activity, pose, and respiration tasks, and it matches or beats strong black-box baselines while remaining interpretable. A subspace regularization scheme, made possible by the transparent architecture, yields an average performance improvement of 19.98% when labeled data are scarce.

What carries the argument

The load-bearing object is the complex sparse rate reduction objective $L(\mathbf{Z}; \mathbf{U}_{[K]}) = R(\mathbf{Z}) - R_c(\mathbf{Z}\mid \mathbf{U}_{[K]}) - \lambda\|\mathbf{Z}\|_0$, where $R$ is a lossy coding rate and $R_c$ measures coding rate within learned subspaces. Because this real-valued objective is not holomorphic in the complex representation matrix $\mathbf{Z}$, the derivations run through CR-Calculus, a differential calculus for real-valued functions of complex variables, used for all gradient and Taylor-expansion steps. The two derived operators are the RF multi-head subspace attention (RF-MSSA), whose core is the complex-softmax attention $U_k^H Z\, S((U_k^H Z)^H (U_k^H Z))$, and the RF-MLP, a residual layer $cReLU(Z + \eta \Omega_Z Z - \eta\lambda \mathbf{1})$ obtained from a proximal majorization-minimization step in which the Hermitian positive-definite metric tensor $\Omega_Z$ is relaxed into a learnable weight. A subspace regularization penalty on per-subspace feature densities ties the interpretable structure to handling scarce labeled data.

What would settle it

A direct ablation would settle it: replace the complex softmax in RF-SSA with the von Neumann-approximated term it is meant to approximate, namely $I - \beta (U_k^H Z)^H (U_k^H Z)$, and retrain on Widar3.0; if accuracy collapses, the softmax is an independently chosen module rather than an output of the derivation. A second check is to re-impose the metric-tensor constraints by rewriting $\Omega_Z = LL^H$ with a learned $L$ and see whether the reported gains persist; if they do not, the derivation is not doing the work claimed.

Watch

Extended reading notes

Core claim

On its own terms, the paper establishes that the entire RF-CRATE architecture — a transformer-like stack of RF self-attention blocks followed by an RF-MLP — can be derived as unrolled iterations of an optimization program, with no black-box block left in the model. The objective is a complex version of sparse rate reduction, and the two alternating minimization steps produce, respectively, a complex multi-head subspace attention operator (query-key-value built from subspace projections with a complex softmax) and a residual feed-forward layer whose skip connection emerges from a proximal majorization-minimization step. Extending CRATE to the complex field is not cosmetic: on the paper's evaluations the complex design adds a 5.08% average classification gain and a 10.34% regression-error reduction over the real-valued predecessor, and the subspace-regularized model matches or exceeds domain-specific baseline models on five datasets across three RF modalities.

Load-bearing premise

The derivation stands on treating the complex softmax as a legitimate stand-in for the subspace-membership term in the gradient step, and on letting the metric tensor become an arbitrary trainable matrix even though the mathematics that produced it required Hermitian positive-definiteness; if either is judged unjustified, the claim that the whole model is mathematically derived is weakened.

Editorial extensions

If this is right

  • If the derivation is accepted, wireless sensing models no longer need to trade interpretability for accuracy: the same unrolled-optimization recipe can be reused for new RF modalities by re-deriving the attention and MLP blocks for the relevant signal statistics.
  • Subspace regularization becomes a general lever for white-box architectures: because the model's subspace occupancy is observable, a simple density-balancing penalty yields an average performance improvement of 19.98% on small RF datasets.
  • Complex-valued processing is validated as more than bookkeeping: the reported gains over the real-valued predecessor (5.08% classification, 10.34% regression error) suggest phase information is being used constructively rather than discarded.
  • The cross-user, cross-environment, and cross-device results from the Widar3.0 experiments imply that a derived, interpretable architecture can generalize across domain shifts without specialized domain-adaptation modules.
  • The subspace correlation, sparsity, and occupancy analyses provide model-health diagnostics that could support unsupervised and semi-supervised RF sensing when labeled data are scarce.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Read charitably, the 'mathematically derived' claim is best understood as an algorithm-unrolling claim: the network structure is the unrolled optimizer, while the softmax replacement and the unconstrained metric tensor are modeling choices inside that unrolling; this reading sets a useful bar for judging future white-box claims.
  • The same complex-softmax attention pattern could transfer to other complex-valued signal domains, such as communications receivers, MRI, and radar imaging, where phase structure is similarly informative; nothing in the derivation is specific to wireless sensing.
  • A natural testable extension is to enforce $\Omega_Z = LL^H$ and compare performance; if accuracy is preserved, the relaxed metric is harmless and the derivation is robust, while a performance drop would show the metric tensor carries capacity the theory did not supply.
  • The subspace-regularization idea may apply to real-valued CRATE-like transformers as well, by monitoring subspace occupancy in the real domain and potentially extending the 19.98% gain beyond RF data.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 6 minor

Summary. The paper proposes RF-CRATE, a complex-valued transformer architecture for radio-frequency (RF) sensing, and claims that the architecture is mathematically derived from a complex sparse rate reduction objective using the CR-Calculus framework. The model consists of an RF self-attention module (RF-MSSA/RF-SSA), an RF-MLP residual module, and a subspace regularization (SSR) technique. The authors evaluate RF-CRATE on five datasets spanning WiFi CSI, FMCW mmWave radar, and IR-UWB, comparing against transformer baselines and specialized DWS models, and report competitive accuracy and regression performance, as well as gains from SSR and from the complex-valued formulation.

Significance. If the derivation claims were fully supported, RF-CRATE would be a notable first result: a fully complex-valued white-box transformer for wireless sensing with interpretable attention and MLP blocks. The paper also provides an open-source implementation, evaluates across heterogeneous RF modalities, and proposes an SSR scheme that appears to improve performance under limited training data. These are real strengths. However, the central theoretical claim is currently overstated: key steps in the derivation—most importantly the replacement of the Gram matrix by a complex softmax in Eq. (14) and the conversion of the metric tensor in Eq. (20) into an unconstrained learnable weight—are design choices rather than mathematical consequences. The paper does not supply machine-checked proofs or approximation bounds for these steps. Because the 'white-box, fully derived' identity is the paper's main selling point, the significance of the work as presented is substantially reduced, even though the empirical study may still be a useful engineering contribution.

major comments (5)
  1. [Section 3.3, Eq. (14), Appendix 9.2] The derivation from the rate-reduction gradient to RF-SSA is not a mathematical consequence: the Gram matrix G_k = (U_k^H Z)^H (U_k^H Z) is replaced by the row-normalized complex softmax S(G_k) without any identity, norm bound, or variational argument connecting the two. The two matrices are qualitatively different: G_k is an unbounded correlation matrix, while S(G_k) is bounded and row-stochastic. Appendix 9.2 states only that 'we can indicate subspace-membership with the complex softmax operation,' which is an architectural choice, not a derived result. In addition, the appendix replaces Sigma_k U_k U_k^H Z by Z in the gradient without stating the needed assumption that Z lies in the span of the U_k's. Consequently, Eq. (12) does not implement a gradient-descent step for Eq. (9), and the claim in Section 3.5 that the entire RF-CRATE model is mathematically derived is unsupported as written.
  2. [Section 3.4, Eqs. (17)-(20)] The metric tensor Omega_Z is required to be Hermitian and positive definite in Eqs. (17) and (19), but Eq. (20) drops these constraints and treats Omega_Z as an arbitrary learned weight matrix. Without those constraints, cReLU(Z + eta Omega_Z Z - eta lambda 1) is a generic residual linear layer, not a step forced by the minimization of Eq. (10). Moreover, componentwise cReLU applied to the real and imaginary parts is not the proximal operator of lambda ||Z||_1 + chi_{Z >= 0} for complex Z; no complex-valued counterpart of Eq. (18) is provided. The RF-MLP module is therefore not actually derived from the stated optimization problem.
  3. [Section 3.4, Eqs. (19)-(20)] The threshold in Eq. (19) is -16 lambda / (9 alpha), while Eq. (20) uses -eta lambda with eta = 16 / (9(1 + alpha)). These two expressions are equal only if alpha = 1 + alpha, which is impossible, so Eq. (20) is not the same update as Eq. (19). Since Eq. (20) defines the deployed RF-MLP operator, this numerical inconsistency must be resolved before the derivation can be considered coherent.
  4. [Section 3.6, Eq. (22), Section 5.4, Fig. 16] SSR is defined as a subspace-density statistic, but the paper never specifies the total training loss that incorporates it, the regularization weight, or how it is combined with the classification or regression loss. Without this information, the reported 19.98% average improvement from SSR cannot be reproduced from the manuscript, and the attribution of the gain to SSR is not testable.
  5. [Section 5.1, Table 1, Fig. 11] The text states that RF-CRATE exhibits 'comparable or superior performance' relative to the highest-performing DWS models, but Table 1 does not support this on all tasks. On mmWave Pose, RF-CRATE reports MPJPE 17.09 versus RF-Net's 8.37, and on UWB Activity it reports 60.07% versus Swin-T's 67.02%. The mean differences are large, and while the confidence intervals overlap, the current wording overstates the empirical case. The claims should be softened to 'competitive on some tasks' or supported by a formal statistical comparison.
minor comments (6)
  1. [Section 3.1] There is a typo in 'from from high-dimensional, complex-valued data'; the repeated word should be removed.
  2. [Section 5.4, Fig. 18] In the sentence 'gesture recognition tast with Widar3.0 dataset', 'tast' should be 'task', and 'MPJE' should read 'MPJPE' in the RF-MLP versus RF-ISTA paragraph.
  3. [Section 3.6, Eq. (22)] The summation index runs from k = 0 to K, whereas subspaces are indexed 1 to K elsewhere; also the notation z_{k,i} and rho_k is introduced only after Eq. (22) and should be defined before use.
  4. [Section 4, patching preprocessing] The CLS token is written as kappa ~ N(0,1) in C^{1x1xd}, but kappa is used both as a random variable and as a token; the notation should clarify that the token is drawn from this distribution and is not itself the scalar kappa.
  5. [Section 5.1, baselines] The description of 'controlled modifications' to baseline models is too vague to assess comparability; the authors should list exactly which hyperparameters or input formats were changed, or point to a documented configuration in the open-source release.
  6. [Throughout] The dataset name is inconsistently spelled as 'Widar3.0' and 'WiDar3.0'; it should be standardized.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular dependency: the unforced derivation choices weaken the 'fully derived' claim but do not reduce any prediction to its inputs.

full rationale

The paper's central object, RF-CRATE, is evaluated against external baselines on public datasets (Table 1), and the reported gains are empirical, so the performance claims are not fitted-then-predicted. The mathematical derivation chain from the complex sparse rate reduction objective (Eq. 5-8) to the update rules (Eq. 11-12 and Eq. 16-19) is presented as algebraic manipulation, and Appendix 9.1 derives the complex rate formula from sphere packing rather than assuming the architecture. The two most fragile steps—the substitution of the complex softmax S(·) for the Gram matrix in Eq. 14 (Appendix 9.2: 'we can indicate subspace-membership with the complex softmax operation') and the relaxation of the Hermitian positive-definite metric tensor Ω_Z to an unconstrained learnable weight in Eq. 20—are not claimed to be identities; they are acknowledged modeling choices. Such gaps undermine the strength of the 'fully mathematically derived' claim, but they do not make the architecture equivalent to its inputs by construction, nor do they rename a fitted parameter as a prediction. The self-citations that exist (e.g., Widar3.0 [75], SLNet [62], and [69]) are dataset, baseline, or motivation citations and are not load-bearing for the derivation. Accordingly, no circular step satisfies the standard of exhibiting a reduction to inputs.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central claim rests on importing CRATE's real-valued rate-reduction lemmas into the complex domain, on several hand-chosen approximations (von Neumann truncation, complex softmax, cReLU, nonnegativity relaxation), and on a learnable metric tensor in RF-MLP. These are not derived from first principles in the paper; they are assumptions or hyperparameters. The empirical evaluation compares against external baselines, so the performance claim is not circular, but the architectural derivation claim is weaker than stated.

free parameters (6)
  • Metric tensor Omega_Z = learned during training
    In RF-MLP (Eq. 20), Omega_Z is treated as an arbitrary learnable weight matrix and the Hermitian positive-definite constraint is relaxed, so the derived update is actually parameterized by fitted weights.
  • Sparsity coefficient lambda = not reported
    Appears in the objective Eq. 5 and in the cReLU threshold in Eq. 19; its value is a hyperparameter that shapes the MLP update.
  • Coding rate constants alpha and beta = not reported
    Defined in Eqs. 6-7; they set the attention scaling and the MLP learning rate eta, but their values are not given in the experiments.
  • Attention gradient step kappa = not reported
    Eq. 12 interpolates between Z and RF-MSSA with kappa; this is a chosen hyperparameter, not derived from the objective.
  • SSR penalty weight = not reported
    Eq. 22 defines the SSR measure but the training loss weight is never specified, even though SSR is credited with a 19.98% average improvement.
  • Number of blocks and subspace parameters K, p, d = model sizes given, but K, p, d not specified
    Architecture hyperparameters; model sizes (7.1M to 80.3M parameters) are given but their correspondence to K, p, d is not stated.
assumptions (5)
  • domain assumption The complex sparse rate reduction objective is a valid measure of representation parsimony for RF data
    Section 3.2 assumes sensing-relevant information is sparse and low-dimensional in complex subspaces; this motivates the objective but is not proven.
  • domain assumption CRATE's Hessian upper-bound lemmas hold in the complex domain for column-normalized Z
    Appendix 9.3 imports results from CRATE [64, P60/P61] to bound the quadratic residual; the paper does not re-derive them for complex inputs.
  • domain assumption The von Neumann series approximation converges for beta (U^H Z)^H (U^H Z)
    Eq. 11 and Appendix 9.2 truncate the inverse series to first order; this requires a spectral radius below 1, which is not checked.
  • ad hoc to paper The complex softmax replacement preserves the subspace-membership association needed by the gradient step
    Appendix 9.2 states 'we can indicate subspace-membership with the complex softmax operation'; this substitutes a softmax of squared moduli for the autocorrelation matrix product, a choice not entailed by the gradient derivation.
  • standard math The l0 norm can be relaxed to l1 plus nonnegativity, and the proximal operator of their sum is cReLU
    Eqs. 15, 18-19 follow CRATE [64] and standard LASSO proximal methods; cReLU is applied elementwise to real and imaginary parts, which is a modeling choice.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Towards White-Box Deep Wireless Sensing." pith.science (2026). https://pith.science/paper/2IQT36IF

@misc{pith2026250721799,
  author       = {Pith},
  title        = {Pith review of: Towards White-Box Deep Wireless Sensing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2IQT36IF}},
  note         = {Machine review of arXiv:2507.21799}
}
read the original abstract

The empirical success of deep learning has spurred its application to the radio-frequency (RF) domain, leading to significant advances in Deep Wireless Sensing (DWS). However, most existing DWS models remain black boxes, with ad-hoc architectures and learned representations lacking explicit physical and mathematical grounding, which limits their reliability and generalizability in real-world deployments. We present RF-CRATE, an early step towards white-box DWS grounded in the complex sparse rate reduction principle. Using the CR-Calculus framework, we derive a fully complex-valued transformer with mathematically interpretable self-attention and residual modules. To address labeled data scarcity, we introduce subspace regularization to enhance representation diversity, yielding a 19.98% average improvement. We evaluate RF-CRATE across heterogeneous RF modalities and human sensing tasks, including activity, gait, and gesture recognition, pose estimation, and respiration monitoring. Experiments on five datasets show that RF-CRATE remains competitive with strong black-box models while providing mathematically interpretable architectures and representations. Moreover, the complex-valued design achieves a 3.39% gain in classification accuracy and a 10.34% reduction in regression error. Our results demonstrate that mathematically grounded models can achieve strong performance in wireless sensing, offering a promising step towards physically aligned white-box DWS systems.

Figures

Figures reproduced from arXiv: 2507.21799 by the authors.

Figure 1
Figure 1. Sparse rate reduction principle. The sparse rate reduction optimization process aims to identify a parsimo￾nious representation space by compressing learned represen￾tations into lower-dimensional subspaces and sparsifying them, thereby enhancing the model’s generalizability. 2.2 White-box Transformer: CRATE Representation learning aims to learn a continuous mapping, 𝑓 (·), that transforms data x to a feature vector… view at source ↗
Figure 2
Figure 2. Overview of RF-CRATE: From first principle to model structure. a significant milestone in deep representation learning. How￾ever, its application has been thus far limited to computer vision tasks, making its extension to RF sensing an exciting yet uncharted problem. Remark: Although CRATE employs the Hermitian transpose symbol in its derivation, it is limited to real-valued inputs. Adapting CRATE requires a theoret… view at source ↗
Figure 3
Figure 3. Parsimony in gesture recognition using WiFi. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: RF Multi-Head Subspace Attention. The imple￾mentation of Eq. 13 serves as the self-attention mechanism specifically designed for RF data in RF-CRATE. RF-MSSA (Multi-Head Subspace Attention) is defined as: RF − MSSA(Z 𝑙 | U 𝑙 [𝐾] ) ≜ 𝛽        U 𝑙⊤ 1 . . . U 𝑙⊤ 𝐾 …
Figure 5
Figure 5. Figure 5: RF-MLP module. The implementation of Eq. 20 acts as the linear layer with a skip connection in the complex field specifically designed for RF data in RF-CRATE model. To minimize −𝑅(Z), we first derive its quadratic upper bound 𝑄(Z | Z 𝑙+1/2 ) in the neighborhood of the…
Figure 6
Figure 6. Figure 6: Subspace correlation heatmaps. Subspace cor￾relation heatmaps showing the degree of uncorrelated sub￾spaces in the feature representations learned by RF-CRATE. (a) (b)              [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 8
Figure 8. Figure 8: Subspace occupancy w.r.t. classification cate [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]
Figure 10
Figure 10. Figure 10: Dataset collection setup. (a) Experiment en [PITH_FULL_IMAGE:figures/full_fig_p009_10.png]
Figure 12
Figure 12. Figure 12: RF-CRATE vs. CRATE on classification tasks. WiP-pose WiP-breath HuPR Datasets 0 2 4 6 8 10 12 14 Error 2.87 2.81 2.44 2.44 2.48 1.76 CRATE RF-CRATE [PITH_FULL_IMAGE:figures/full_fig_p011_12.png]
Figure 14
Figure 14. Figure 14: Performance evalua￾tion of RF-CRATE across multiple users and environments. Rx1 Rx2 Rx3 Tx Rx4 Rx5 Rx6 50 60 70 80 90 86.7 75.3 81.8 65.8 84.7 74.0 77.3 80.9 76.5 77.9 77.8 User orientation Cross-orientation Cross-device Tx: loc of transmitter Rx: loc of source receiv…
Figure 17
Figure 17. Figure 17: RF-CRATE perfor￾mances with raw signal input or DFS spectrum. zReLU cardioid modReLU cReLU 50 60 70 80 90 100 Accuracy (%) [PITH_FULL_IMAGE:figures/full_fig_p013_17.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. RF-LEGO: Modularized Signal Processing-Deep Learning Co-Design for RF Sensing via Deep Unrolling

    cs.DC 2026-04 unverdicted novelty 6.0 of 10

    RF-LEGO turns signal processing algorithms into trainable modular DL modules via deep unrolling, outperforming pure SP and DL baselines in RF sensing while preserving interpretability.

Reference graph

Works this paper leans on

78 extracted references · 44 canonical work pages · cited by 1 Pith paper

  1. [1]

    Samira Abnar and Willem Zuidema. 2020. Quantifying Attention Flow in Transformers. https://doi.org/10.48550/arXiv.2005.00928 arXiv:2005.00928 [cs]

    arXiv doi:10.48550/arxiv.2005.00928 2020

  2. [2]

    Fadel Adib, Zach Kabelac, Dina Katabi, and Robert C. Miller. 2014. 3D Tracking via Body Radio Reflections. In 11th USENIX Symposium on Networked Systems Design and Implementation (NSDI 14) . 317–329

    work page 2014

  3. [3]

    Martin Arjovsky, Amar Shah, and Yoshua Bengio. 2016. Unitary Evo- lution Recurrent Neural Networks. arXiv:1511.06464 [cs, stat]

    arXiv 2016

  4. [4]

    Sebastian Bach, Alexander Binder, Grégoire Montavon, Frederick Klauschen, Klaus-Robert Müller, and Wojciech Samek. 2015. On Pixel-Wise Explanations for Non-Linear Classifier Decisions by Layer- Wise Relevance Propagation. PLOS ONE 10, 7 (2015), e0130140. https://doi.org/10.1371/journal.pone.0130140

    doi:10.1371/journal.pone.0130140 2015

  5. [5]

    Kaifeng Bi, Lingxi Xie, Hengheng Zhang, Xin Chen, Xiaotao Gu, and Qi Tian. 2023. Accurate Medium-Range Global Weather Forecasting with 3D Neural Networks. Nature 619, 7970 (July 2023), 533–538. https://doi.org/10.1038/s41586-023-06185-3

    doi:10.1038/s41586-023-06185-3 2023

  6. [6]

    Bocus, Wenda Li, Shelly Vishwakarma, Roget Kou, Chong Tang, Karl Woodbridge, Ian Craddock, Ryan McConville, Raul Santos-Rodriguez, Kevin Chetty, and Robert Piechocki

    Mohammud J. Bocus, Wenda Li, Shelly Vishwakarma, Roget Kou, Chong Tang, Karl Woodbridge, Ian Craddock, Ryan McConville, Raul Santos-Rodriguez, Kevin Chetty, and Robert Piechocki. 2022. OPER- Anet, a Multimodal Activity Recognition Dataset Acquired from Radio Frequency and Vision-Based Sensors. Scientific Data 9, 1 (Dec. 2022),

    work page 2022

  7. [7]

    David H Brandwood. 1983. A Complex Gradient Operator and Its Ap- plication in Adaptive Array Theory. In IEE Proceedings H (Microwaves, Optics and Antennas), Vol. 130. IET Digital Library, 11–16

    work page 1983

  8. [8]

    Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. 2021. Emerging Prop- erties in Self-Supervised Vision Transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision . 9650–9660

    work page 2021

Show all 78 references
  1. [9]

    Rao, Troy D

    Supriyo Chakraborty, Richard Tomsett, Ramya Raghavendra, Daniel Harborne, Moustafa Alzantot, Federico Cerutti, Mani Srivastava, Alun Preece, Simon Julier, Raghuveer M. Rao, Troy D. Kelley, Dave Braines, Murat Sensoy, Christopher J. Willis, and Prudhvi Gurram. 2017. In- terpret...

    2017

  2. [10]

    Kwan Ho Ryan Chan, Yaodong Yu, Chong You, Haozhi Qi, John Wright, and Yi Ma. 2022. ReduNet: A White-box Deep Network from the Principle of Maximizing Rate Reduction. Journal of Machine Learning Research 23, 114 (2022), 1–103

    2022

  3. [11]

    Chaofan Chen, Oscar Li, Daniel Tao, Alina Barnett, Cynthia Rudin, and Jonathan K Su. 2019. This Looks Like That: Deep Learning for Interpretable Image Recognition. In Advances in Neural Information Processing Systems, Vol. 32. Curran Associates, Inc

    2019

  4. [12]

    Xi Chen, Yan Duan, Rein Houthooft, John Schulman, Ilya Sutskever, and Pieter Abbeel. 2016. InfoGAN: Interpretable Representation Learning by Information Maximizing Generative Adversarial Nets. https://doi.org/10.48550/arXiv.1606.03657 arXiv:1606.03657 [cs, stat]

    2016 doi arXiv

  5. [13]

    Yu Chen, Ying Tai, Xiaoming Liu, Chunhua Shen, and Jian Yang. 2018. Fsrnet: End-to-end learning face super-resolution with facial priors. In Proceedings of the IEEE conference on computer vision and pattern recognition. 2492–2501

    2018

  6. [14]

    Zhe Chen, Tianyue Zheng, Chao Cai, and Jun Luo. 2021. MoVi-Fi: Motion-Robust Vital Signs Waveform Recovery via Deep Interpreted RF Sensing. In Proceedings of the 27th Annual International Conference on Mobile Computing and Networking (MobiCom ’21) . Association for Computing M...

    2021

  7. [15]

    Guoxuan Chi, Zheng Yang, Chenshu Wu, Jingao Xu, Yuchong Gao, Yunhao Liu, and Tony Xiao Han. 2024. RF-Diffusion: Radio Signal Generation via Time-Frequency Diffusion. In Proceedings of the 30th Annual International Conference on Mobile Computing and Networking . 77–92

    2024

  8. [16]

    Junyoung Chung, Caglar Gulcehre, KyungHyun Cho, and Yoshua Bengio. 2014. Empirical evaluation of gated recurrent neural networks on sequence modeling. arXiv preprint arXiv:1412.3555 (2014)

    2014 arXiv

  9. [17]

    Shuya Ding, Zhe Chen, Tianyue Zheng, and Jun Luo. 2020. RF-net: A Unified Meta-Learning Framework for RF-enabled One-Shot Human Activity Recognition. InProceedings of the 18th Conference on Embedded Networked Sensor Systems (SenSys ’20) . Association for Computing Machinery, N...

    2020

  10. [18]

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weis- senborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. 2021. An Image Is Worth 16x16 Words: Transformers for Image Recogn...

    2021 doi arXiv

  11. [19]

    Alexey Dosovitskiy and Thomas Brox. 2016. Inverting Visual Repre- sentations With Convolutional Networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition . 4829–4837

    2016

  12. [20]

    Florian Eilers and Xiaoyi Jiang. 2023. Building Blocks for a Complex- Valued Transformer Architecture. In ICASSP 2023 - 2023 IEEE Interna- tional Conference on Acoustics, Speech and Signal Processing (ICASSP) . 1–5. https://doi.org/10.1109/ICASSP49357.2023.10095349 arXiv:2306....

    2023

  13. [21]

    Alhussein Fawzi, Matej Balog, Aja Huang, Thomas Hubert, Bernardino Romera-Paredes, Mohammadamin Barekatain, Alexander Novikov, Francisco J R Ruiz, Julian Schrittwieser, Grzegorz Swirszcz, et al. 2022. Discovering faster matrix multiplication algorithms with reinforce- ment lea...

    2022

  14. [22]

    Nitzan Guberman. 2016. On Complex Valued Convolutional Neural Networks. arXiv:1602.09046 [cs]

    2016 arXiv

  15. [23]

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2016. Deep Residual Learning for Image Recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition . 770–778

    2016

  16. [24]

    Yuze He, Chen Bian, Jingfei Xia, Shuyao Shi, Zhenyu Yan, Qun Song, and Guoliang Xing. 2023. VI-Map: Infrastructure-Assisted Real-Time HD Mapping for Autonomous Driving. In Proceedings of the 29th An- nual International Conference on Mobile Computing and Networking (ACM MobiCom...

    2023

  17. [25]

    Akira Hirose. 2006. Complex-valued neural networks. Springer

    2006

  18. [26]

    Wein- berger

    Gao Huang, Zhuang Liu, Laurens van der Maaten, and Kilian Q. Wein- berger. 2018. Densely Connected Convolutional Networks. https: //doi.org/10.48550/arXiv.1608.06993 arXiv:1608.06993 [cs]

    2018 doi arXiv

  19. [28]

    John Jumper, Richard Evans, Alexander Pritzel, Tim Green, Michael Figurnov, Olaf Ronneberger, Kathryn Tunyasuvunakool, Russ Bates, Augustin Žídek, Anna Potapenko, Alex Bridgland, Clemens Meyer, Simon A. A. Kohl, Andrew J. Ballard, Andrew Cowie, Bernardino Romera-Paredes, Stani...

    2025

  20. [29]

    Sangwon Kim, Jaeyeal Nam, and Byoung Chul Ko. 2022. ViT-NeT: Interpretable Vision Transformers with Neural Tree Decoder. In Pro- ceedings of the 39th International Conference on Machine Learning . PMLR, 11162–11172

    2022

  21. [30]

    Ken Kreutz-Delgado. 2009. The Complex Gradient Operator and the CR-Calculus. arXiv:0906.4835 [math]

    2009 arXiv

  22. [31]

    Shankar Krishnan, Ying Xiao, and Rif A Saurous. 2017. Neumann optimizer: A practical optimization algorithm for deep neural networks. arXiv preprint arXiv:1712.03298 (2017)

    2017 arXiv

  23. [32]

    Shih-Po Lee, Niraj Prakash Kini, Wen-Hsiao Peng, Ching-Wen Ma, and Jenq-Neng Hwang. 2023. HuPR: A Benchmark for Human Pose Esti- mation Using Millimeter Wave Radar. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (W ACV). 5715– 5724

    2023

  24. [33]

    Gupta, and Dezhi Hong

    Shuheng Li, Ranak Roy Chowdhury, Jingbo Shang, Rajesh K. Gupta, and Dezhi Hong. 2021. UniTS: Short-Time Fourier Inspired Neural Networks for Sensory Time Series Classification. In Proceedings of the 19th ACM Conference on Embedded Networked Sensor Systems (SenSys ’21). Associa...

    2021

  25. [34]

    Xuhong Li, Haoyi Xiong, Xingjian Li, Xuanyu Wu, Xiao Zhang, Ji Liu, Jiang Bian, and Dejing Dou. 2022. Interpretable Deep Learn- ing: Interpretation, Interpretability, Trustworthiness, and Beyond. Knowledge and Information Systems 64, 12 (Dec. 2022), 3197–3234. https://doi.org/...

    2022 doi

  26. [35]

    Xiang Li, Daqing Zhang, Qin Lv, Jie Xiong, Shengjie Li, Yue Zhang, and Hong Mei. 2017. IndoTrack: Device-free indoor human tracking with commodity Wi-Fi. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies 1, 3 (2017), 1–22

    2017

  27. [36]

    Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, Furu Wei, and Baining Guo. 2022. Swin Transformer V2: Scaling Up Capacity and Resolution. https://doi.org/10.48550/arXiv.2111.09883 arXiv:2111.09883 [cs]

    2022 doi arXiv

  28. [37]

    Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. 2021. Swin Transformer: Hierarchical Vision Transformer Using Shifted Windows. https://doi.org/10.48550 /arXiv.2103.14030 arXiv:2103.14030 [cs]

    2021 doi arXiv

  29. [38]

    Ilya Loshchilov and Frank Hutter. 2019. Decoupled Weight Decay Regularization. arXiv:1711.05101 [cs.LG] https://arxiv.org/abs/1711.0 5101

    2019 arXiv

  30. [39]

    Yi Ma, Harm Derksen, Wei Hong, and John Wright. 2007. Segmentation of Multivariate Mixed Data via Lossy Data Coding and Compression. IEEE Transactions on Pattern Analysis and Machine Intelligence 29, 9 (Sept. 2007), 1546–1562. https://doi.org/10.1109/TPAMI.2007.1085

    2007

  31. [40]

    Vishal Monga, Yuelong Li, and Yonina C. Eldar. 2020. Algorithm Unrolling: Interpretable, Efficient Deep Learning for Signal and Image Processing. arXiv:1912.10557 [cs, eess]

    2020 arXiv

  32. [41]

    Nils Mönning and Suresh Manandhar. 2018. Evaluation of complex- valued neural networks on real-valued classification tasks. arXiv preprint arXiv:1811.12351 (2018)

    2018 arXiv

  33. [42]

    Chris Olah, Alexander Mordvintsev, and Ludwig Schubert. 2017. Fea- ture Visualization. Distill 2, 11 (Nov. 2017), e7. https://doi.org/10.239 15/distill.00007

    2017

  34. [43]

    Qifan Pu, Sidhant Gupta, Shyamnath Gollakota, and Shwetak Patel

  35. [44]

    Kun Qian, Chenshu Wu, Yi Zhang, Guidong Zhang, Zheng Yang, and Yunhao Liu. 2018. Widar2.0: Passive Human Tracking with a Single Wi-Fi Link. In Proceedings of the 16th Annual International Conference on Mobile Systems, Applications, and Services (MobiSys ’18). Association for C...

    2018

  36. [45]

    Sara Sabour, Nicholas Frosst, and Geoffrey E. Hinton. 2017. Dynamic Routing Between Capsules. Advances in Neural Information Processing Systems 30 (2017), 3856–3866

    2017

  37. [46]

    Simone Scardapane, Steven Van Vaerenbergh, Amir Hussain, and Aurelio Uncini. 2020. Complex-Valued Neural Networks With Non- parametric Activation Functions. IEEE Transactions on Emerging Top- ics in Computational Intelligence 4, 2 (April 2020), 140–150. https: //doi.org/10.110...

    2020

  38. [48]

    Ruiyuan Song, Dongheng Zhang, Zhi Wu, Cong Yu, Chunyang Xie, Shuai Yang, Yang Hu, and Yan Chen. 2022. RF-URL: Unsupervised Representation Learning for RF Sensing. In Proceedings of the 28th Annual International Conference on Mobile Computing And Networking (MobiCom ’22). Assoc...

    2022

  39. [49]

    Mohammad Tofighi, Tiantong Guo, Jairam KP Vanamala, and Vishal Monga. 2019. Prior information guided regularized deep learning for cell nucleus detection. IEEE transactions on medical imaging 38, 9 (2019), 2047–2058

    2019

  40. [50]

    Chiheb Trabelsi, Olexa Bilaniuk, Ying Zhang, Dmitriy Serdyuk, Sandeep Subramanian, João Felipe Santos, Soroush Mehri, Negar Rostamzadeh, Yoshua Bengio, and Christopher J. Pal. 2018. Deep Complex Networks. https://doi.org/10.48550/arXiv.1705.09792 arXiv:1705.09792 [cs]

    2018 doi arXiv

  41. [51]

    David Tse and Pramod Viswanath. 2005. Fundamentals of Wireless Communication (1 ed.). Cambridge University Press. https://doi.org/ 10.1017/CBO9780511807213

    2005 doi

  42. [52]

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. At- tention Is All You Need. In Advances in Neural Information Processing Systems, Vol. 30. Curran Associates, Inc

    2017

  43. [53]

    Yu, and Michael Lustig

    Patrick Virtue, Stella X. Yu, and Michael Lustig. 2017. Better than Real: Complex-valued Neural Nets for MRI Fingerprinting. In 2017 IEEE International Conference on Image Processing (ICIP) . 3953–3957. https://doi.org/10.1109/ICIP.2017.8297024

    2017

  44. [54]

    Yu, and Michael Lustig

    Patrick Virtue, Stella X. Yu, and Michael Lustig. 2017. Better than Real: Complex-valued Neural Nets for MRI Fingerprinting. https: //doi.org/10.48550/arXiv.1707.00070 arXiv:1707.00070 [cs]

    2017 doi arXiv

  45. [55]

    Yan Wang, Jian Liu, Yingying Chen, Marco Gruteser, Jie Yang, and Hongbo Liu. 2014. E-Eyes: Device-Free Location-Oriented Activity Identification Using Fine-Grained WiFi Signatures. In Proceedings of the 20th Annual International Conference on Mobile Computing and Networking (M...

    2014

  46. [56]

    Chenshu Wu, Feng Zhang, Yusen Fan, and K. J. Ray Liu. 2019. RF-based Inertial Measurement. In Proceedings of the ACM Special Interest Group on Data Communication (SIGCOMM ’19) . Association for Computing Machinery, New York, NY, USA, 117–129. https://doi.org/10.1145/33 41302.3342081

    2019

  47. [57]

    Chenshu Wu, Feng Zhang, Yuqian Hu, and K. J. Ray Liu. 2021. GaitWay: Monitoring and Recognizing Gait Speed Through the Walls. IEEE 16 RF-CRATE Arxiv, 2025 Transactions on Mobile Computing 20, 6 (June 2021), 2186–2199. https: //doi.org/10.1109/TMC.2020.2975158

    2021

  48. [58]

    Jin-Hui Wu, Shao-Qun Zhang, Yuan Jiang, and Zhi-Hua Zhou. 2023. Complex-Valued Neurons Can Learn More but Slower than Real-valued Neurons via Gradient Descent. In Advances in Neural Information Processing Systems, Vol. 36. 23714–23747

    2023

  49. [59]

    Tianfu Wu and Xi Song. 2019. Towards Interpretable Object Detection by Unfolding Latent Structures. In 2019 IEEE/CVF International Confer- ence on Computer Vision (ICCV). IEEE, Seoul, Korea (South), 6032–6042. https://doi.org/10.1109/ICCV.2019.00613

    2019

  50. [60]

    C. Xiao, D. Han, Y. Ma, and Z. Qin. 2019. CsiGAN: Robust Channel State Information-Based Activity Recognition With GANs. IEEE Internet of Things Journal 6, 6 (Dec. 2019), 10191–10204. https://doi.org/10.1109/ JIOT.2019.2936580

    2019

  51. [61]

    Jinrui Yang, Xianhang Li, Druv Pai, Yuyin Zhou, Yi Ma, Yaodong Yu, and Cihang Xie. 2024. Scaling White-Box Transformers for Vision. arXiv preprint arXiv:2405.20299 (2024)

    2024 arXiv

  52. [62]

    Zheng Yang, Yi Zhang, Kun Qian, and Chenshu Wu. 2023. {SLNet}: A Spectrogram Learning Neural Network for Deep Wireless Sensing. In 20th USENIX Symposium on Networked Systems Design and Implemen- tation (NSDI 23). 1221–1236

    2023

  53. [63]

    Shuochao Yao, Ailing Piao, Wenjun Jiang, Yiran Zhao, Huajie Shao, Shengzhong Liu, Dongxin Liu, Jinyang Li, Tianshi Wang, Shaohan Hu, Lu Su, Jiawei Han, and Tarek Abdelzaher. 2019. STFNets: Learning Sensing Signals from the Time-Frequency Perspective with Short- Time Fourier Ne...

    2019

  54. [64]

    Haeffele, and Yi Ma

    Yaodong Yu, Sam Buchanan, Druv Pai, Tianzhe Chu, Ziyang Wu, Shengbang Tong, Hao Bai, Yuexiang Zhai, Benjamin D. Haeffele, and Yi Ma. 2024. White-Box Transformers via Sparse Rate Reduction: Com- pression Is All There Is? Journal of Machine Learning Research (2024). https://doi....

    2024 doi arXiv

  55. [65]

    Pathak, and Prasant Mohapatra

    Yunze Zeng, Parth H. Pathak, and Prasant Mohapatra. 2016. WiWho: Wifi-Based Person Identification in Smart Spaces. In Proceedings of the 15th International Conference on Information Processing in Sensor Networks (IPSN ’16). IEEE Press, Vienna, Austria, 1–12

    2016

  56. [66]

    Jie Zhang, Zhanyong Tang, Meng Li, Dingyi Fang, Petteri Nurmi, and Zheng Wang. 2018. CrossSense: Towards Cross-Site and Large-Scale WiFi Sensing. InProceedings of the 24th Annual International Conference on Mobile Computing and Networking (MobiCom ’18) . Association for Comput...

    2018

  57. [67]

    Quanshi Zhang, Ying Nian Wu, and Song-Chun Zhu. 2018. Inter- pretable Convolutional Neural Networks. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition . IEEE, Salt Lake City, UT, 8827–8836. https://doi.org/10.1109/CVPR.2018.00920

    2018

  58. [68]

    Quan-shi Zhang and Song-chun Zhu. 2018. Visual Interpretability for Deep Learning: A Survey. Frontiers of Information Technology & Electronic Engineering 19, 1 (Jan. 2018), 27–39. https://doi.org/10.163 1/FITEE.1700808

    2018

  59. [69]

    Xie Zhang, Chengpei Tang, Kang Yin, and Qingqian Ni. 2021. WiFi- based Cross-Domain Gesture Recognition via Modified Prototypical Networks. IEEE Internet of Things Journal (2021), 1–1. https://doi.org/ 10.1109/JIOT.2021.3114309

    2021

  60. [70]

    Yi Zhang, Yue Zheng, Guidong Zhang, Kun Qian, Chen Qian, and Zheng Yang. 2020. GaitID: Robust Wi-Fi Based Gait Recognition. In Wireless Algorithms, Systems, and Applications (Lecture Notes in Computer Science), Dongxiao Yu, Falko Dressler, and Jiguo Yu (Eds.). Springer Interna...

    2020

  61. [71]

    Zizhao Zhang, Han Zhang, Long Zhao, Ting Chen, Sercan Ö Arik, and Tomas Pfister. 2022. Nested Hierarchical Transformer: Towards Accurate, Data-Efficient and Interpretable Visual Understanding. Pro- ceedings of the AAAI Conference on Artificial Intelligence 36, 3 (June 2022), 3...

    2022 doi

  62. [72]

    Mingmin Zhao, Tianhong Li, Mohammad Abu Alsheikh, Yonglong Tian, Hang Zhao, Antonio Torralba, and Dina Katabi. 2018. Through- wall human pose estimation using radio signals. In Proceedings of the IEEE conference on computer vision and pattern recognition . 7356–7365

    2018

  63. [73]

    Mingmin Zhao, Yonglong Tian, Hang Zhao, Mohammad Abu Alsheikh, Tianhong Li, Rumen Hristov, Zachary Kabelac, Dina Katabi, and Anto- nio Torralba. 2018. RF-based 3D skeletons. In Proceedings of the 2018 Conference of the ACM Special Interest Group on Data Communication . 267–281

    2018

  64. [74]

    Tianyue Zheng, Zhe Chen, Shujie Zhang, Chao Cai, and Jun Luo. 2021. MoRe-Fi: Motion-robust and Fine-grained Respiration Monitoring via Deep-Learning UWB Radar. In Proceedings of the 19th ACM Conference on Embedded Networked Sensor Systems (SenSys ’21) . Association for Computi...

    2021

  65. [75]

    Yue Zheng, Yi Zhang, Kun Qian, Guidong Zhang, Yunhao Liu, Chen- shu Wu, and Zheng Yang. 2019. Zero-Effort Cross-Domain Gesture Recognition with Wi-Fi. In Proceedings of the 17th Annual International Conference on Mobile Systems, Applications, and Services (MobiSys ’19) . Assoc...

    2019

  66. [76]

    1 𝑚 𝑚∑︁ 𝑖=1 ˆw𝑖 ˆw𝐻 𝑖 # . Then, substituting the expression for ˆw𝑖, we get: ˆΣ = E

    Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. 2015. Object Detectors Emerge in Deep Scene CNNs. https: //doi.org/10.48550/arXiv.1412.6856 arXiv:1412.6856 [cs] 17 Arxiv, 2025 RF-CRATE 9 APPENDIX 9.1 Complex lossy coding rate For the sake of compl...

    2015 doi arXiv

  67. [80]

    Note, here we use Z to replace Zt to avoid confusion and use the conjugate coordinates Z, Z for𝑅

    Upper bound∥HZZ(Z𝑡)∥ℓ2→ℓ2: Based on the definition of the operator norm, we have ∥HZZ(Z𝑡)∥ℓ2→ℓ2 ≜ sup ∥Δ∥≤ 1 ∥HZZ(Z𝑡)Δ∥𝐹 with Δ∈C 𝑑×𝑁 To calculate the Hessian matrix, we use the method in [64] [𝐶𝑅𝐴𝑇𝐸,𝑃 59, 127]: if Δ is any matrix with the same shape as 𝑍 and𝑡 > 0, HZZ(Z)( Δ) ...

    2025

  68. [81]

    To this end, we have the quadratic upper bound of−𝑅(Z): −𝑅(Z)≤− 𝑅(Z𝑙+1/2)+ D −∇z𝑅(Z𝑙+1/2), Z− Z𝑙+1/2 E + 1 2 ∫ 1 0 (1−𝑡) 𝛼+𝛼 8 Z− Z𝑙+1/2 2 𝐹 𝑑𝑡

    Upper bound HZZ(Z𝑡) ℓ2→ℓ2 : Follow the same method above, we have HZZ(Z𝑡) ℓ2→ℓ2 ≜ sup ∥Δ∥≤ 1 HZZ(Z𝑡)Δ 𝐹 = sup ∥Δ∥≤ 1 𝜕 𝜕𝑡′ 𝑡′=0 h 𝑡′↦→∇ Z𝑡𝑅 Z𝑡, Z𝑡+𝑡′Δ i 𝐹 = sup ∥Δ∥≤ 1 𝛼 2 Z −𝛼 I+𝛼Z𝐻 𝑡 Z𝑡 −1 Δ⊤Z𝑡 I+𝛼Z𝐻 𝑡 Z𝑡 −1 𝐹 = 𝛼 2 2 sup ∥Δ∥≤ 1 Z𝛼 I+𝛼Z𝐻 𝑡 Z𝑡 −1 Δ⊤Z𝑡 I+𝛼Z𝐻 𝑡 Z𝑡 −1 𝐹 ≤ 𝛼 2 2 ...

    2025

  69. [474]

    https://doi.org/10.1038/s41597-022-01573-2

    doi

  70. [2013]

    In Proceedings of the 19th Annual International Conference on Mobile Com- puting & Networking

    Whole-Home Gesture Recognition Using Wireless Signals. In Proceedings of the 19th Annual International Conference on Mobile Com- puting & Networking. 27–38

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.